Decision support for ARMA model identification using hierarchically organized neural networks

To resolve the difficulties in autoregressive moving average (ARMA) model identification, the extended sample autocorrelation function (ESACF) is adopted as a feature extractor, and the multilayered backpropagation network (MLBPN) is used as a pattern classifier. To improve the classification power of MLBPNs, a hierarchically organized neural network is proposed, which consists of an AR network and many small-sized MA networks. The output of the AR network determines the AR order of a time series, and designates the MA network which will give the MA order. A step-by-step training strategy is also suggested so that the learned MPBPNs can effectively classify ESACF patterns contaminated by a high level of noise. The experiment with the artificially generated test data and real world data showed promising results